Identification of Emotion Parameters in Music to Modulate Human Affective States
Maíra Araújo de Santana, Ingrid Bruno Nunes, Flávio S. Fonseca, Arianne S. Torcate, Amanda Suarez, Vanessa Marques, Nathália Córdula, Juliana Carneiro Gomes, Giselle Machado Magalhães Moreno, Wellington Pinheiro dos Santos · 2023
Emotions are a complex aspect of human being that are being widely explored to improve human-machine interaction. This complexity makes emotions the object of study in several works, particularly the ones that aim to further understand human emotional behavior or to use affective states to control interfaces and/or create personalized content. In this context, emotion recognition is one of the key topics. The effectiveness of this recognition depends on the way the different emotions are being represented. There are two main ways to categorize emotions: the discrete model, in which emotions are directly labeled; and the two-dimensional model, which associates emotions to secondary parameters such as valence and arousal. The dimensional approach has particularly drawn attention since it can provide a more fine-grained emotion assessment. In this approach, each emotion may be represented as function of valence and arousal continuous values. However, the automatic identification of valence and arousal degrees from a particular content is a difficult task, which may be benefited by the use of artificial intelligence algorithms. Therefore, this chapter proposes an approach to the automatic identification of emotions induced by music of different genres based on the prediction of their valence and arousal parameters. To achieve this goal we extracted explicit numerical features from the songs and used this set of features to train regression models based on Linear Regression, SVM, ELM, Random Forest and MLP. Random Forest model outperformed the others with better correlation coefficients and lower errors for predicting both arousal and valence values from the music signals. Arousal prediction with Random Forest resulted in a Pearson’s correlation coefficient of 0.85 – 0.01. While valence was predicted with a correlation of 0.76 – 0.02.